Your AI assistant forgets everything when you close the session. For your whole team, that's a productivity tax you pay every single day. Synapse fixes it — permanently, collectively, with audit trails.
55 MCP tools, 258 skills, multi-user authentication, 12-page dashboard, collaborative whiteboards, team-safe knowledge graph — served from a single self-hosted server. Claude Code, Claude Desktop, or any MCP client connects and gets the full system.
Self-improving AI brain for Claude Code & Desktop — 28 MCP tools, 253 skills, collective memory, project tracking, work logs. One server, all your sessions share the same knowledge. Deploy on Coolify in 2 minutes.
Restart Claude Code. You now have 50 tools, 258 skills, and shared team memory.
First login creates the admin user automatically. Additional team members are invited from the dashboard (Team page) and each gets their own API key for MCP access.
For Claude Desktop, add the same to ~/Library/Application Support/Claude/claude_desktop_config.json.
See It In Action
Resume a project after days away
> session_resume({ project: "Quickfy" })
## Resume Context: Quickfy
### Last Session
- Date: 2026-04-15
- Task: add stripe payments
- Status: paused
- Summary: Implemented Stripe checkout + webhook handler
- Next steps: Add subscription renewals and account page
- Branch: feat/payments
- Commits: abc123, def456
### Project Memories (3)
- [BugFix conf:3] Cookie forwarding in Server Actions needs explicit headers()
- [Pattern conf:3] Centralize Payload Local API calls in lib/payload.ts
> memory_query({ scope: "team", type: "BugFix" })
M-bugfix-xxx [BugFix conf:3] @alice • 2026-04-20
"Every collection in multi-tenant Payload MUST have access control functions..."
M-bugfix-yyy [BugFix conf:2] @bob • 2026-04-22
"Coolify healthchecks fail on slow migrations — increase start_period to 90s"
A bug Alice fixed last week is now searchable from Bob's session. Personal notes stay private (scope: "personal"), project-specific knowledge scopes to the project, and team learnings are visible to everyone.
Self-service API keys
Each team member rotates their own MCP access token from the Profile page. No shared passwords, no server restart, full audit trail.
What's New — May 2026
Whiteboards (FigJam-style, integrated with Synapse data) — full collaborative whiteboard with infinite canvas, sticky notes, frames, shapes, code blocks, free-hand pen, images, emoji stamps, and connectors. The killer feature: drag Synapse memories/skills/sessions/projects directly onto the canvas as live cards, double-click to edit them inline, and run 14 deterministic generators to auto-build boards from your data (by-tag, by-type, by-project, by-skill, by-author, recent, most-used, decisions-log, antipatterns, open-todos, session-timeline, skill-graph, project-overview, semantic-cluster). Includes async collaboration (comments + @mention email + in-app notification inbox + heartbeat presence avatars), 5 templates (Retro, Brainstorm, Kanban, OKR, Mind-map), undo/redo, snap-to-grid + alignment guides, multi-handle resize, group/ungroup, lock, dot voting, timer, color-by-author, markdown sticky, syntax-highlighted code, version snapshots auto-saved hourly, soft-delete trash, pin/favorites, tags + search, read-only public share links, full board export (PNG/JPG/PDF/JSON), MCP exposure (5 new tools), and Figma-style left vertical sidebar. Zero npm dependencies (vanilla JS, ~3000 LOC), MIT-only vendor libs (marked, Prism, html2canvas) self-hosted in /vendor/.
Semantic search (vector embeddings) — memory_search now uses hybrid retrieval: 0.5 × BM25 + 0.5 × cosine similarity via Xenova/multilingual-e5-small (384-dim, CPU-only, IT+EN). Memories that share meaning surface even without keyword overlap.
Embed-on-add — every new memory is embedded in the background automatically (fire-and-forget, never blocks memory_add). First call downloads the model (~118 MB) to the local cache.
Graceful degradation — if the model is unavailable, retrieval falls back to BM25-only. No breaking changes.
Backfill script — pnpm backfill:embeddings re-embeds all existing memories in batches of 32 (idempotent, safe to re-run after deploy).
What's New — April 2026
The last wave of releases turned Synapse from a single-user tool into a team platform. Everything below ships in the current build:
Multi-user authentication — email + password, admin bootstrap on first login, legacy DASHBOARD_PASSWORD deprecated.
Self-service API keys — GET/POST/DELETE /api/me/api-keys from the Profile page, no admin needed for rotation.
Memory scopes — personal / team / project on every memory. "My memories" filter isolates what you authored.
Team-safe storage layer — ownership tracking (created_by_user_id, updated_by_user_id) across memories, skills and components, with ConcurrencyError on stale writes.
Review system & approval queue — memories and skills can be saved as draft and approved from the Review page. Full audit log (review_audit table) with approve/reject/rollback.
Skill version history — every upsert snapshots the skill into skill_versions with a change_reason (manual, import, haiku-evolution, rollback). Instant rollback to any prior version.
Design system auto-populate — scan Tailwind v4 configs, CSS variables and tokens.json files. Cross-referenced against the component catalog.
Component catalog + token linking — bulk import component snippets, auto-detect var(--color-*) references, link to design systems.
Project merge & dedup — POST /api/projects-meta/merge absorbs duplicate projects (sessions, memories, env vars, design systems merge with conflict resolution).
MemPalace retrieval — BM25 re-ranking on FTS5 trigram candidates (partial/prefix match: "serv" → "server"), closet boost for project-context results, verbatim session chunk indexing with session_search.
Memory persistence fix — CodeGraph uploads (codegraph_analyze) now merge only the index tables; memories, sessions, skills and env vars are never overwritten.
ENCRYPTION_KEY rotation — POST /api/admin/rotate-key re-encrypts all secrets atomically; safe to run before changing the key in Coolify.
CodeGraph auto-upload — proxy auto-uploads the local analysis on connect if the server is stale or missing; Bearer auth fixed for /api/* routes.
Before vs After
Without Synapse
With Synapse
Same bug, 3rd time
20 min rediscovering
memory_search → 2 seconds
New session
Start from zero
Auto-loads context + top memories
Multiple sessions
Isolated silos
One shared database — instant sync
Resume after days
"Where was I?"
session_resume → exact state
"Don't do X"
Forgotten next session
AntiPattern with decay — persists
Stale knowledge
Old patterns pollute
Auto-decay: unused → deprecated
Which skill to use?
Search docs manually
skill_route → instant match
Team collaboration
Isolated inboxes / Slack
Shared DB, team scope, per-user ownership
Access control
Trust everyone
Per-user API keys + audit log
Knowledge curation
Everything auto-merged
draft status + approval queue
Brainstorm / planning meeting
Miro/FigJam disconnected from your data
Whiteboard with 14 generators that pull memories/decisions/sessions live
Visual project handoff
Slides + 5 stale wikis
Project overview board · share read-only link · auto-extract memories from retro
What Is Synapse?
A self-improving AI brain deployed on your server, with multi-user authentication and team-safe storage. Thirteen integrated systems:
📚 1. Skills-as-a-Service (258 skills)
Domain knowledge served via MCP: Next.js, Stripe, SEO, CRO, Payload CMS, Docker, React Native, design systems, and more. Loaded on demand.
skill_list() → browse all 258 skills
skill_route(task) → find the best skills for your task
skill_read(name) → load full skill content
skill_add(name) → contribute a new skill (draft mode supported)
skill_update(name) → update a skill, history kept in skill_versions
🧠 2. Memory Graph
Typed knowledge graph with 8 memory types, 5 relationship types, and 3 scopes, stored in SQLite. Retrieval uses hybrid scoring (0.5 × BM25 + 0.5 × cosine similarity) via multilingual-e5-small (384-dim, CPU-only, IT+EN) over FTS5 trigram candidates (partial/prefix match: "serv" → "server"), plus closet boost for project-context relevance. Falls back to BM25-only when the model is unavailable.
Memory types: Fact, Preference, Decision, Pattern, AntiPattern, BugFix, Goal, Todo
Edge types: RelatedTo, Updates, Contradicts, CausedBy, PartOf
Scopes:personal (private to author), team (visible to everyone), project (scoped to one project)
Every memory has confidence scoring (1–10) with automatic decay. Bad memories fade. Good ones strengthen.
🌐 3. Collective Memory
One shared SQLite database on the server. All clients — Claude Code, Claude Desktop, any MCP client — read and write through the same store. A bug fixed in session A is instantly searchable from session B.
Alice (Claude Code) ──proxy──→ server ──→ SQLite (shared)
Bob (Claude Desktop) ──proxy──→ server ──→ SQLite (shared)
Carol (Browser) ──HTTPS──→ server ──→ Dashboard
👥 4. Multi-User + Teams (NEW)
Email + password authentication with bootstrap admin on first login. Admins invite team members from the Team page, assign roles, and track membership. Every write records the author.
🔑 5. Self-Service API Keys (NEW)
Each user rotates their own MCP access tokens from the Profile page (GET/POST/DELETE /api/me/api-keys). Admin never sees your key. Tokens are hashed at rest.
✅ 6. Review System & Approval Queue (NEW)
Save memories and skills as draft — they land in the Review page queue instead of going live. Admins (or whoever has the review role) approve, reject, or roll back. Every decision is logged in review_audit.
🗂 7. Skill Version History (NEW)
Every time a skill is created or updated, a snapshot goes into skill_versions with a change_reason (manual / import / haiku-evolution / rollback). One click restores any prior version.
🎨 8. Design System Registry (NEW)
Scan Tailwind v4 configs, CSS variables and tokens.json files directly — the MCP accepts the file contents inline. The design system is stored, merged, and cross-referenced with the component catalog.
🧩 9. Component Catalog (NEW)
Register reusable UI components (hero, pricing table, CTA block, etc.) with code snippet, section type, and design-token references auto-detected from var(--color-*) usage. Searchable by project, section, or token.
📁 10. Project Tracking + Work Log + Merge
Projects are auto-derived from sessions. Every session records task, status, deliverables, next steps, branch, and commits. Duplicates can be merged without losing history.
project_list() → all projects with status
project_merge({ from, into }) → absorb a duplicate project
session_resume({ project }) → full context to continue work
📊 11. Hub Dashboard
Web dashboard served on the same port as the MCP HTTP server. Bearer-token auth shared with MCP. Twelve pages:
Home — stats, recent memories, sessions
Projects — all projects with status and context
Work Log — deliverables timeline per project
Skills — browse/search 258 skills, edit & version
Memories — explore memories with scope / type / project / author filters
Sessions — timeline of all sessions across the team
Components — component catalog with design-token linking
Design Systems — scan, merge and inspect design tokens
Team — invite users, assign roles, manage members
Review — approval queue for draft memories and skills
Profile — rotate your API keys, change password
⚡ 12. Auto-Session Tracking
The proxy auto-detects your project (from package.json or folder name), git branch, and client type (Code vs Desktop). Sessions are created, heart-beated and closed automatically.
🎨 13. Whiteboards (NEW — May 2026)
A FigJam-style infinite canvas, but wired to your Synapse data. Drag any memory / skill / session / project onto the board as a live card and edit it in place. Run 14 generators to auto-build boards from your knowledge graph (by tag, by type, by project, by skill, by author, recent, most-used, decisions log, antipatterns, open todos, session timeline, skill graph, project overview, semantic cluster via K-means + embeddings). Sticky notes with markdown, code blocks with syntax highlighting, frames, free-hand pen, shapes (rect/ellipse/triangle), images (drag-drop, base64 inline), emoji stamps, connectors with 4 kinds (related/depends-on/blocks/leads-to). Full async collaboration: comments threading, @email mentions (email + in-app inbox), heartbeat-based presence avatars, dot voting, timer, color-by-author. 5 templates (Retro, Brainstorm 3-step, Kanban, OKR, Mind-map). Auto-snapshots hourly, soft-delete trash + restore, pin/favorites, tags, search across boards, read-only public share links, export PNG/JPG/PDF/JSON. Zero JS dependencies (vanilla, ~3000 LOC); only 3 MIT vendor libs self-hosted (marked, Prism, html2canvas).
whiteboard_list({ scope, projectName, tag, pinned, search }) → boards
whiteboard_create({ name, scope, nodes, tags, description }) → new board
whiteboard_add_nodes({ id, nodes, connectors }) → append
whiteboard_read({ id }) → full state
whiteboard_search({ q }) → matches
Dashboard preview
55 MCP Tools
Memory (8)
Tool
Purpose
memory_add
Save a memory (auto-detects contradictions, supports scope and draft)
memory_search
Hybrid search (BM25 + vector cosine) across all memory fields
memory_query
Filter by type, project, scope, author, skill, confidence
memory_load
Load top-scored memories for current session
memory_add_edge
Create relationships between memories
memory_stats
Statistics, contradictions and active drafts
memory_decay
Apply decay cycle (reinforce / decay / deprecate)
memory_suggest
AI-suggested memories to promote or merge
Skills (11)
Tool
Purpose
skill_list
List all skills with filters
skill_read
Read full skill content
skill_route
Find best skills for a task
skill_stats
Skills statistics by type / category
skill_add
Create a new skill (draft mode supported)
skill_update
Update a skill (versioned automatically)
agent_list
List all registered agents
agent_read
Read agent prompt
command_list
List all slash commands
command_read
Read command content
cortex_briefing
5-layer working memory briefing
Sessions (6)
Tool
Purpose
session_start
Log session with project + task
session_heartbeat
Keep long-running sessions alive
session_end
Close with deliverables + next steps
session_resume
Full context to continue a project
session_history
Recent sessions, filter by project or author
session_search
Full-text search across verbatim past session content
Projects (10)
Tool
Purpose
project_list
All projects with status and context
project_list_full
Full project export including env, stack, CI/CD
project_get
Read a single project by slug
project_scan
Detect stack, CMS, deploy target from repo
project_update
Update project metadata
project_merge
Merge a duplicate project into another
project_set_env
Encrypt & store a single env var
project_set_env_batch
Encrypt & store multiple env vars
project_get_env
Read decrypted env vars for an authorised user
project_generate_env_example
Emit a redacted .env.example
Code Intelligence (7)
Tool
Purpose
codegraph_query
Semantic search by concept
codegraph_context
360-degree view of a symbol
codegraph_impact
Blast radius before editing
codegraph_detect_changes
Map git diff to affected symbols
codegraph_rename
Graph-aware multi-file rename
codegraph_list_repos
All indexed repositories
codegraph_cypher
Raw SQL against the graph
Components & Design Systems (8)
Tool
Purpose
component_add
Register a component with code snippet + tokens
component_list
List all components with filters
component_get
Get a single component
component_search
Full-text search across components
component_scan
Bulk import from file contents
design_system_scan
Parse Tailwind config / CSS vars / tokens.json
design_system_get
Read the design system for a project
design_system_set
Overwrite the design system manually
Whiteboards (5) (NEW)
Tool
Purpose
whiteboard_list
List boards with filters (scope, project, tag, pinned, search)
whiteboard_read
Read a board (metadata + nodes + connectors)
whiteboard_create
Create a new board (with optional seed nodes/tags/description)
Option A: Connect to an existing server (2 minutes)
Get an API key from your admin (or generate one from your Profile page), then add the MCP config as shown in 2-Minute Setup.
Option B: Deploy your own server
1. Clone and build
git clone https://github.com/deve1993/skillbrain
cd skillbrain/packages/codegraph
npm install && npm run build
2. Import skills
node dist/cli.js import-skills ../..
3. Run locally
# Local MCP (stdio — single machine, no auth)
node dist/cli.js mcp
# HTTP server (for remote access + dashboard + multi-user)
node dist/cli.js mcp --http --port 3737
4. Bootstrap the admin user
On first HTTP start, set ADMIN_EMAIL and DASHBOARD_PASSWORD to bootstrap the admin account. After first login these become optional — the admin manages users from the dashboard.
DASHBOARD_PASSWORD is kept for backwards compatibility (migration 012_deprecate_dashboard_password.sql), but the preferred auth path is per-user email + password + API keys.
Q: Does this work with Cursor / Windsurf / OpenCode?
A: Yes — any tool that supports MCP over stdio, through the included HTTP proxy.
Q: Can I have multiple users with different permissions?
A: Yes. Email + password login, admin / user roles, per-user API keys, and every write records the author.
Q: Can I review memories and skills before they go live?
A: Yes. Save them as draft — they land in the Review page queue. Approvals, rejections and rollbacks are logged in review_audit.
Q: What happens to memories created by someone who leaves the team?
A: Ownership is tracked (created_by_user_id). Admins can list and re-assign.
Q: How do I migrate from v1 (shared DASHBOARD_PASSWORD) to v2 (multi-user)?
A: Set ADMIN_EMAIL before first login — the current DASHBOARD_PASSWORD bootstraps the admin account. Generate API keys per team member from the Team page. The old shared password keeps working as a fallback until you remove it.
Q: What happens offline?
A: The proxy needs internet to reach the server. For offline use, run node cli.js mcp directly — local SQLite, no server, no multi-user.
Q: How does it get smarter over time?
A: Every session captures memories (bugs, patterns, decisions). Confidence grows with validation, decays without use. Draft memories sit in the review queue until approved. Skills auto-evolve when Haiku proposes updates that an admin approves.
Q: What inspired this?
A: Spacebot.sh by Spacedrive — their typed memory system with graph edges. We adapted it for MCP + SQLite + self-hosted deployment, and extended it with multi-user auth, approval workflow and skill versioning.
Contributing
Contributions welcome:
New domain skills (submit as draft — admin will review)
Memory Graph integrations
Dashboard improvements
Multi-language support
Migration scripts between Synapse instances
License
Synapse is dual-licensed:
AGPL-3.0 — Free for open source use. If you run a modified version over a network (including SaaS), you must share your source code under AGPL-3.0.
Commercial License — For proprietary/closed-source deployments. Removes all AGPL obligations.
The complete AGPL-3.0 text is in the LICENSE file.
Built with Claude Code + Synapse collective memory
S
Synapse
PIXARTSeu·11 ดาว·
Self-improving AI brain for Claude Code & Desktop — 28 MCP tools, 253 skills, collective memory, project tracking, work logs. One server, all your sessions share the same knowledge. Deploy on Coolify
Self-improving AI brain for Claude Code & Desktop — 28 MCP tools, 253 skills, collective memory, project tracking, work logs. One server, all your sessions share the same knowledge. Deploy on Coolify in 2 minutes.
ตรวจสอบล่าสุดเมื่อ 2026-10-28 Self-improving AI brain for Claude Code & Desktop — 28 MCP tools, 253 skills, collective memory, project tracking, work logs. One server, all your sessions share the same knowledge. Deploy on Coolify in 2 minutes.